Re: [Qemu-devel] [PATCH 00/22] QAPI Round 1

[Anthony Liguori <anthony@codemonkey.ws>]

On 03/08/2011 08:17 AM, Avi Kivity wrote:
> On 03/08/2011 04:10 PM, Anthony Liguori wrote:
>> On 03/08/2011 08:00 AM, Avi Kivity wrote:
>>> On 03/08/2011 03:54 PM, Anthony Liguori wrote:
>>>>
>>>>> (and gah, do we really need a vfs/rpc in qemu?)
>>>>
>>>> Fun, eh :-)  Unfortunately, our friends at VMware provide a 
>>>> VixVM_CopyFileFromGuestToHost API so there's an expectation that we 
>>>> provide a similar interface.
>>>>
>>>
>>> Yes, but do we have to terminate it in qemu?
>>
>> No, I'm in the process of writing up my latest proposal.
>>
>> The idea is pretty simple.  QAPI generates code for libqmp that takes 
>> native arguments for a command and generates a QObject.  It also 
>> generates code for QEMU that takes a QObject and generates native 
>> arguments to pass to a function.
>>
>> For guest commands, we combine the two such that we unmarshal the 
>> incoming QObject to native arguments, then pass it to another 
>> function that marshals the arguments to a QObject.  The QObject is 
>> then passed to the guest-agent which uses the same generated code as 
>> QEMU to unmarshal the qobject to native arguments and dispatch to a 
>> function.
>>
>> That means the only new code we need for the guest agent is the 
>> JSON-over-virtio-serial transport.  To implement guest commands, we 
>> just add the command to the schema, implement the native arguments 
>> version in guest-agent, and that's it.
>>
>> QEMU will buffer all input and output to the guest acting as a first 
>> line of defence from a security PoV.  That means that the guest 
>> doesn't get to talk directly to the management tools which removes 
>> that as a direct attack surface.
>>
>> The nature of QEMU is such that if we do tagging correctly, we can 
>> also support live migration transparently to the guest too.
>
> Okay, do I understand correctly that qemu does not understand each 
> command individually?  It just reads the schema and converts from one 
> rpc protocol to another (even if they are the same protocol)?
>
> So: mgmt -> json -> qemu -> (qobject -> ) json -> guest (and back again).

Exactly.

>
> If that's the case, I like it.

QEMU will have the ability to execute commands to the guest agent and 
that will be used by something like HMP to provide friendly versions of 
the interfaces.  However, my long term plan is to move HMP out of core 
QEMU and the QAPI refactoring is making the HMP code not depend on any 
core QEMU functions, only on the QMP interface.

Regards,

Anthony Liguori